NEW Can a random forest algorithm distinguish between albums?

One of the biggest question I had about my corpus is: Are albums actually distinguishable? To answer this question, I used a random forest algorithm to predict for each track to which album it belongs. In this graph you can see the result of this process. Along the x-axis you see the the actual albums. If you look for example at the first album ‘Bricolage’, you can look across the y-axis to see which albums are predicted the most for tracks from this album. In the case of ‘Bricolage’ the algorithm could correctly predict about half of the tracks, but this is not the case with every album. It is interesting for example that the algorithm flips the two albums ‘Fear in a Handful of Dust’ and ‘Long Stories’, and listening to them you hear that indeed they are very similar. Overall you can conclude that the algorithm does not do very well except for a few exceptions. This may be because the sample size is very small and therefore there are lots of options for overfitting or underfitting.

Introduction and pre-analysis of the corpus

The corpus I’m going to be inspecting is a collection of all albums from the electronic artist Amon Tobin. I chose this artist because he has produced a lot of different albums, ranging from dreamy ambient music (Fear in a Handful of Dust) to very intense beats (Fight! Fight! Fight!). There are interesting comparisons to be made in this corpus. Because his earliest album is from 1997, which is a relatively early year for electronic music, the change in music over time is an obvious comparison to make. Because the corpus consists of only albums, it is also interesting to see how and if albums differ, especially very similar albums like Fear in a Handful of Dust and Long Stories. I expect to find a lot of differences between the ambient albums and the heavier electronic albums, and I expect certain albums to be very difficult to differentiate. A strength of the corpus is that most of the music of Amon Tobin is represented in these albums. A weakness might be that certain albums might be too similar to find any meaningful differences. If expect Fear in a Handful of Dust and Long Stories to be too similar because they both use long dreamy chords and melodies without using drums, which might be difficult for Spotify to differentiate. An atypical track for this corpus might be ‘New York Editor’, for its jazzy drums and basslines.

Full albums used in the corpus:
Bricolage (1997)
Permutation (1998)
Supermodified (2000)
Out From Out Where (2002)
Chaos Theory (2005)
Foley Room (2007)
ISAM (2011)
Electronic Music for the Sydney Opera House (2017)
Fear in a Handful of Dust (2019)
Time To Run (2019)
Long Stories (2019)
Fight! Fight! Fight! (2020)
The World As We Know It (2020)

Are there even large differences in Amon Tobin’s tracks?

Here we see a lot of Amon Tobin’s tracks are high in energy and low in acousticness. An interesting find though is that the more acoustic tracks tend to be less energetic than the less acoustic tracks according to Spotify, which is odd since acoustic albums like Time To Run feel much more energetic than mainly electronic albums like Fear in a Handful of Dust. Valence seems to be about evenly spread, but the few tracks with the lowest valence tend to have very low acousticness.

Are there clear tempo differences between albums?

A spectogram analysis of the outlier track “Big Furry Head”

Do his albums differ in terms of high level Spotify features?

Here you can see there are clear differences between albums. There is also an interesting relationship between energy, valence and danceability, namely albums with high valence tend to have high energy as well as high danceability. Another thing to note is that Amon Tobin’s albums are relatively low in valence, while energy seems to vary much more. Mouse over the plot to see the specific values.

Taking a look at ‘Toys’ through cepstrograms

SSM of the track ‘Reanimator’

This is a self-similarity matrix for timbre of the track ‘Reanimator’, since this track is a very typical track of Amon Tobin according to the Spotify features. There are dark patches here and there which indicate similar timbre throughout the track. Interesting points are the yellow line at around 185 seconds which indicates the moment the heavy drums kick in and the cluster of yellow lines at around 250 seconds which indicate a transitional moment in the track.

tempo and loudness of albums in chronological order

Here you can see the tempo and loudness of all the albums in chronological order. If you look at tempo you see a clear progression in albums from 1997 till 2020, you can sort of draw a curve for tempo in time. It is also interesting that later albums differ much more in loudness than earlier albums. The deviation of tempo (indicated by the boxes) is overall also much higher in later albums. This indicated a clear change in music.

Lets look at the harmonic progression over time

I want to look at how Amon Tobin’s music changed harmonically over the years. To avoid cluttering I’ve split all his tracks around the midpoint of his career, 2007. In this graph you can see the distribution of keys he used in his tracks, as well as the distribution of major and minor tracks. The most interesting find in this graph is the key C#. After 2007 there are more tracks produced in C# minor, but significantly less in C# major. After investigating further, I noticed most of the tracks in C# minor before 2007 were from the album ‘Supermodified’. In the next page I take a closer look at one of those tracks for any interesting harmonic information.

Taking a closer look at the harmonics of multiple C# minor tracks from before 2000